(solution) Journal of Experimental Psychology: Human Learning and

(solution) Journal of Experimental Psychology: Human Learning and

Here attached is an article on Arthur Reber’s 1976 experiment. Can you please explain the theory behind the experiment, the hypothesis (including independent and dependent variable) (the rationale behind the experiment), the procedure, what information did Reber found, the results and why did it happen?

Journal of Experimental Psychology:
Human Learning and Memory
1976, Vol. 2, No. 1, 88-94 Implicit Learning of Synthetic Languages:
The Role of Instructional Set
Arthur S. Reber
Brooklyn College
The effect of instructional set on implicit learning of a synthetic language
was explored. Specifically, the neutral, implicit instructions used in previous studies were compared with explicit instructions which directed subjects to search for the complex rules that determined letter orderings. The
subjects given the explicit instructions were poorer at memorizing exemplars
from the language, learned less about the underlying structure, and tended
to invent nonrepresentational rules. The results have strong implications
for a theory of implicit learning which stresses a nonconscious abstraction
system that operates when the stimulus environment exhibits exceedingly
complex structure, and subjects are not actively trying to break the code. Elsewhere implicit learning has been characterized as a process whereby a subject becomes sensitive to the structure inherent in
in a complex array by developing (implicitly) a conceptual model which reflects
the structure to some degree (Reber, 1967,
1969; Reber & Millward, 1968). It has
been a working hypothesis that the learning
process is fundamentally an abstraction of
information from the environment by the
subject without recourse to explicit strategies for responding or explicit systems for
encoding the stimuli. Support for the hypothesis has been indirect, consisting essentially of the failure of several groups of
researchers to find evidence of verbalizable,
explicit hypothesis testing or rule formation
even when subjects display highly efficient
exploitation of the patterns present in the
stimuli (Braine, 1963; Foss, 1968; Reber,
1967, 1969; Smith & Braine, in press).
Given the highly abstract nature of the
implicit learning process the evidence will
necessarily continue to be indirect. However, such evidence has a way of accumulating until a threshold of legitimacy is
achieved for the hypothesized abstraction.
This research was supported in part by grant
MH-20239-01 from the National Institute of Mental Health.
Requests for reprints should be sent to Arthur S.
Reber, Department of Psychology, Brooklyn College of CUNY, Brooklyn, New York 11210. This study provides an additional quantum
of data.
One of the obligations incurred whenever
a process is hypotheized is the specification
of the boundary conditions, the circumstances under which the presumed process
will occur. One limiting condition in implicit learning clearly must be the lower
bound on the complexity of the underlying
structure that defines the stimuli. That is,
the stimulus patterns must be such that they
bias against the possibility of there being
appropriate coding schemes available to the
subjects. The issue is essentially one of definition : If the subject can discover and
formalize the rule system that characterizes
the stimulus array then the experiment is
no longer an experiment in implicit learning; it becomes one in inductive rule learning or rule identification. It was this consideration that prompted the highly complex
structures in the past and dictates the similarly complex system in this study.
What is not so obvious, however, is the
question of the "set" of the subject when
confronted with these complex arrays. In
previous work great care has been taken to
ensure that subjects were neutral with regard to underlying structure, and thus presumably neutral with respect to the use of
explicit hypothesis-testing strategies. This
study was designed to explore the impact
that the operating set has upon the subject's
performance. The procedure used was IMPLICIT LEARNING OF SYNTHETIC LANGUAGE straightforward: The behavior of subjects
run under neutral instructions was compared with that of subjects who were given
general information about synthetic grammars and encouraged to undertake an explicit search for rules. More specifically,
the aim was to explore the differences between (a) subjects who maintained a relatively naive stance with regard to rule structure and operated in a neutral mode insofar
as the formulation of hypotheses and strategies is concerned, and (b) subjects who
actively searched for rules and operated in
an explicit hypothesis-testing mode.
METHOD
Stimulus Items
As in previous studies the stimulus items were
strings of letters generated by a finite-state grammar (Figure 1). This grammar may be characterized as a Markovian process in which each
permissible transition from any one state, Si, to
another state, Sj, generates a symbol. A grammatical string in the language is defined as any
sequence of permissible transitions leading from
the initial state, So, to the terminal state,So'. The
language is defined as all possible paths through
the system. For example, the sequence of states
So-Sa-Sa-Si-Sz-Ss-Si-So' generates the acceptable
sequence VXVPXVS. This particular grammar
generates exactly 43 permissible letter sequences
of lengths three through eight which were used as
the grammatical stimulus materials for the experiment. (See Chomsky & Miller, 1958, and Reber,
1967, for details of the procedures involved in these
calculations.) Subjects
The subjects were 20 undergraduates who served
as part of a course requirement. They were randomly assigned to the two groups. Procedure
The experiment was run in two parts, a learning phase and a testing phase. Prior to the beginning of the learning phase, subjects were read
the instructions appropriate for the group they
were in. Except for these instructions all subjects were run identically.
The instructions for the implicit group (Group
I) were as follows:
This is a. simple memory experiment. You will
see items made up of the letters PSTVX. They
will run from three to eight letters in length
and will be shown to you in groups of three
items each. After seeing each set of three items
I will give you a card and your task will be to 89 FIGURE 1. Schematic diagram of the finite-state
grammar used to generate the stimuli. (So = initial
state; So' = terminal state. The language is all
possible paths through the system.)
try to reproduce all three items. I will tell
you which ones you reproduced correctly. After
you have reproduced all three correctly two
times in a row we will go on to a new set of
three items.
The instructions for the explicit group (Group
E) were the same as above with the addition of
the following:
The order of letters in each item is determined
by a rather complex set of rules. The rules only
allow certain letters to follow other letters. Since
the task involves memorization of a large number of these complex strings of letters, it will be
to your advantage if you can figure out what the
rules are, which letters may follow other letters
and which ones may not. Such knowledge will
certainly help you in learning and memorizing
the items.
Learning. Fifteen of the 43 grammatical strings
were selected as stimulus items for this phase of the
experiment. They were selected as representative
of the possible types of grammatical strings; for
example, for each length, strings beginning with
both T and V were included, and for all lengths
where it was possible, an example of each of the
loops of the grammar (P, X, VPX) was used.
These 15 items were presented to subjects in five
sets of 3 items each. Each item was printed on a
separate card and presented through a viewing window for 5 sec. After the 3 items of a given set
were shown, the subject was given a card and
asked to reproduce all 3 items. There were no
time restrictions on the subjects although long response times were rare. Subjects were told which
items were correctly reproduced but no information
was given about the nature of the errors. Each
set was run repeatedly in the same order until the
criterion of two consecutive correct reproductions
was reached, after which the next set was presented. The order of presentation was varied randomly for each subject. All subjects continued
until all 15 items were learned. A S-min rest ARTHUR S. REBER 90 TABLE 1
MEAN ERRORS TO CRITERION BY GROUP
Set
M Group I
E 8.3
17.4 4.9
9.5 4.0
7.7 2.0
5.3 2.3
4.5 4.03
8.88 Note. Abbreviations: I = implicit instructions; E = explicit
instructions. period was allowed before the beginning of the
test phase.
Testing. Twenty-two of the remaining grammatical items were selected along with 22 items
that violated the rules of the grammar. Four of
these nongrammatical items were formed randomly
and contained multiple violations; the remaining
18 contained only single-letter violations.
The stimulus items were presented through the
same viewing window. The subjects' task was to
make a decision about the correctness or grammaticality of each item based upon what they had
learned during the initial memorization phase and
to then press one of two buttons marked "yes"
and "no." Note that none of the subjects had been
told at the outset that there would be a testing
phase, and that for the subjects in Group I this
was the first time that any reference to rules had
been made.
The list of 44 test items was presented twice,
making a total of 88 for each subject. All subjects were informed about the equal proportions of
grammatical and nongrammatical items. There was
no time limit on the subjects during this phase
although they were told that latencies were being
recorded. No feedback about the correctness of a
decision was given until the full set of 88 items
was completed. RESULTS1 is high. Overall, Type 1 strings 2 were
somewhat easier to learn than other types,
and this tendency was observed in both
groups. The other four types were equally
difficult although Group E subjects experienced uniformly more difficulty than Group
I subjects and made more error on all types.
Testing
Note first that the lack of feedback about
the correctness of a response served to keep
subjects at a constant level of performance
throughout the testing phase. Also, as in
previous work, subjects were generally not
aware of the fact that each test item was
presented twice. Thus in the following
analyses the full set of 88 test items is
treated as a single block. The data are
presented in Table 2.
TABLE 2
FREQUENCY OF G AND NG RESPONSES TO G
AND NG ITEMS IN BOTH GROUPS
Item
Response NG Total Group I
G
NG 334
106 99
341 433
447 Group E
G
NG 263
177 131
309 394
486 Learning Note. Group differences are highly significant, p < .001; correct responses for both groups are significantly better than
chance, fa < .001. Abbreviations; G = grammatical; NG
= nongrammatical; I = implicit instructions; E = explicit
instructions. The learning phase data are most easily
expressed in terms of errors to criterion. As
Table 1 shows there was a strong difference
between the groups with those given explicit
instructions performing significantly poorer,
*(18) = 4.75, SEAltt = 2.14. Trials to criterion data were comparable and are not
presented.
In general the learning data were similar
to those found in earlier work (Reber, 1967,
1969). Subjects in both groups eventually
adopted the procedure of focusing upon only
one or two strings per memorization trial,
a common strategy when information load 1
The rejection region throughout is p < .01
unless otherwise noted.
2
The finite-state grammar in Figure 1 generates
five basic sentence types. Each type is defined by
a path through the system with obligatory and
optional transitions, the latter being the loops
or recursions. The five types are as follows:
(1) T[P]TS; (2) T[P]TX[X](VPX[X])VS;
(3) T[P]TX[X](VPX[X])VPS; (4) V[X]
(VPX[X])VS; and (5) V[X] (VPX[X])VPS.
Note that Types 2 and 3 and Types 4 and 5 are
very similar differing only in the obligatory P in
the next-to-last position in Types 3 and 5; Type 1
appears, superficially, to be considerably simpler
than the other four. IMPLICIT LEARNING OF SYNTHETIC LANGUAGE
Both groups showed the fruits of the
learning phase and were able to discriminate
grammaticality (G) from nongrammaticality
(NG) at far better than chance levels, fs(9)
= 22.3 and 15.2 for Groups I and E, respectively, SEM = .59 and 1.54, respectively.
Moreover, every one of the 20 subjects
showed discriminability above chance.
_ However, the two groups differed from each
other considerably in this ability, with Group
E subjects being poorer than those given
the implicit instructions, <(18) = 6.24, SEAUt
= 1.74. Group I performance agreed nicely
with earlier findings and, indeed, is actually
a replication of work reported in Reber
(1967); Group E performance fell far below this level. Note also that Group E
shows a strong bias toward nongrammatical
responses. The argument is made later in
this article that this bias can be expected
under the assumption that the primary effect
of the specific instructions to these subjects
was to produce a tendency for them to develop rule systems which were not representative of the underlying structure.
A variety of other, more fine-grain,
analyses were carried out, none of which
provided any important insight into the
qualitative differences between the two
groups, and all of which were comparable to
similar results found in Reber (1967). For
example, the nongrammatical items with the
violation in either the initial or the terminal
letter were detected better than items with
the violation in internal positions. Items
which contained multiple violations were
easier to detect as nongrammatical than
items with single letter violations. The five
item types were all equally likely to be recognized as grammatical. Naturally, the overall level of performance in all these cases
was lower for Group E subjects but the
general pattern was the same in both groups.
Further, the latency data failed to reveal
group differences. Both groups had shorter
latency distributions for items on which
grammaticality was correctly assessed than
on items where an error was made, p < .01
for Group I and p < .02 for Group E (Kolmogorov-Smirnov test). There were no differences in response time to assign grammaticality as opposed to nongrammaticality 91 for either group, and there were no betweengroups differences.
The one statistic that does reveal an important quantitative difference between the
mode of operation of the two groups is the
probability of a subject making an error on
both presentations of a given item (Pe,e)
compared with the probability of an error
on only one of the two presentations (Pe,c
and Pc,e). In terms of a simple detection
model, Pc,c = fc + (1-*) g2
Pc,e = P e , c = ( l – £ ) < 7 ( l – 0 ) P… = (!-*) ( l – < 7 ) 2 , where k is a parameter which reflects the
subjects' level of apprehension of the grammatical relations, that is, the probability of
knowing the grammaticality of any given
item. The probability of a correct guess
is represented as g, and by virtue of the
equal proportions of G and NG items, g =
.50. The known value of P<.tC was used to
estimate k and thus "predict" the other
values.
In principle, k can be estimated from any
of the equations, although Pc,c is the appropriate source since this value alone is
assumed to contain instances where the subjects knew the grammatical status of the
letter strings. Further, estimating k in this
manner provides for a cleaner test of the
strong prediction of the model, that is,
Pc,e = Pe,e = Pe,e> which IS 3 direct result of fixing g at .50.
Interpreting the model is straightforward:
It is deemed appropriate only under the
condition where the subjects' decisions are
based upon an accurate (although partial)
representation of the rules of the grammar.
Inappropriate representative rules will create
instances where subjects incorrectly assume
that they know the grammatical status of
test items and will produce an inflated value
of Pe,e relative to Pc,e and Pe,c. Thus the
model serves as a sensitive test of the implicit nature of the subjects' behavior as it
pertains to the error data.
In evaluating the model x2 tests for goodness-of-fit were carried out with the values
of PeiC and Pc,e pooled. The pooling was
done so that any perturbations produced by 92 ARTHUR S. REBER
TABLE 3
TEST OF THE SIMPLE DETECTION MODEL
Probability
Result Group I
Predicted
Obtained
Group E
Predicted
Obtained .66
.66 .11
.10 .11
.11 .11
.13 .53
.53 .16
.12 .16
.12 .16
.23 vaue o
= .55 and .37 for Groups
I and E, respectively.
model is rejected for
Group E. x!(10) = 44.06, p < .001, bu
not for Group I, x2(10) = 8.86. fluctuations in these values would not contribute to x2, and thus any rejection of the
model must be due to the inflated value
of Pe,e- These fluctuations were, of course,
nonsystematic, as can be seen from the
group averages given in Table 3 where Pe,c
and Pc,e are essentially identical for both
groups.
Each individual subject was tested against
the model and the individual x2 values were
summed to produce the group results shown
in Table 3. Using p < .05 as the critical
value, 7 of the 10 subjects in Group E had
values of Pe,e in excess of what would be
expected by chance alone, while only 1 such
subject was found in Group I. For completeness, all predicted and obtained values
are given in Table 3, although the x2 tests
were all run with the pooling described
before.
The overall effects are quite clear. The
model is clearly an inappropriate characterization of the behavior of Group E subjects
while for Group I it is well within the
expected range. Subjects who operate in
an implicit mode develop abstract representations which are accurate (if partial) reflections of the structure of the stimulus
items. Subjects who are instructed to perform explicit rule searches develop abstractions which also contain representations
that are inaccurate reflections of the underlying structure. It is also worth noting that,
independent of the appropriateness of the
rule systems developed, both groups are
equally consistent in applications. Summing the values of .Pc,c and Pe,e yields .79 for
Group I and .76 for Group E, so that in
some sense each group has learned the same
"number" of rules.
Careful subject-by-subject analyses were
carried out in an effort to pinpoint qualitative
differences, particularly on those items responded to consistently. Occasionally systematicity was discovered, such as one E
subject who (erroneously) rejected any
item with a letter repeated more than four
times (despite the fact that TPPPPPTS
was among the set of learning items), and
another who accepted (also erroneously)
any item where the P-cycle was misplaced
(e.g., TTPPPXVS). These cases, however, were relatively rare and account for
a trivial amount of the large group differences that were found. The rich and
complex rule system which, as is argued
later, is necessary to produce implict learning, carries with it the serious liability that
subjects' abstractions will be similarly rich,
complex, and intractable.
DISCUSSION
There are several straightforward conclusions to be drawn from these data: (a) Subjects who engaged in an explicit search for
rules that define a complex structure performed more poorly in memorizing exemplars of the structure than subjects who
operated in a more neutral, implicit fashion,
(b) Although taken to the same learning
criterion, subjects operating in the explicit
mode ultimately learned less about the
underlying properties of the complex stimuli
than subjects in the implicit mode, (c) Explicit search for rules produced a strong
tendency for subjects to induce or invent
rules which were not accurate representations of the complex stimulus structure; this
tendency was not observed in subjects given
the implicit instructions.
The critical word in each of the foregoing
conclusions is complex. Except for it, these
conclusions would be at variance with the
rule-learning literature in which instructions that focus the subject's attention upon
the rule system generally accelerate learning.
In these other typically explicit rule-learning
experiments, since the stimulus patterns are IMPLICIT LEARNING OF SYNTHETIC LANGUAGE relatively simple and codable, a subject with
a reasonably rich stock of heuristics and
problem-solving strategies is going to find
their implementation rewarded. For example, essentially all of the work in serial
pattern learning (see Jones, 1974) has been
carried out using stimulus sequences whose
underlying structures can be coded by a subject equipped with search strategies based
upon devices like alternating events, eventrun length, complementation, and so forth.
However powerful these strategies may
be, the structure of the strings of letters
generated by the grammar in Figure 1 is
simply too rich to be coded by a subject
using them in the short time allowed. The
implication is that the explicit instructions
disrupted performance by inviting, indeed
encouraging, subjects to engage in futile
rule-search procedures and to elaborate rule
systems which were frequently nonrepresentational. The slower learning rate, the bias
toward assigning nongrammaticality, and the
inflated value of Pe,e all support this interpretation. It is important to note that the
instructions to Group E are not specifically
misleading. They interfere with performance, not because they mislead the subjects, but because they put them in an
operating set where they mislead themselves.
For example, one not atypical subject exhibited consternation during the postexperimental debriefing when she discovered that
her elaborate and sophisticated efforts to
find remote, deterministic contingencies between letters were doomed to fail. It is
trivially true, then, that searching for rules
will not work unless you can find them.
It should be emphasized here that explicit
rule search can jeopardize its user in
another, perhaps more significant, way. In
addition to producing poor performance because of a failure to find well-formed rules,
engaging in explicit rule search acts to mask
the implicit learning process. As has been
suggested elsewhere (Reber, 1967; Reber,
& Millward, 1968), the implicit acquisition
process seems to be most effective when the
subjects are in a relatively neutral, passive
set and allow themselves to be inundated by
the stimulus materials. The efforts on the
part of Group E subjects to break the code 93. precludes the operation of this implicit mode.
A serious difficulty with these experiments
on implicit learning is determining just how
they blend in with the traditional work on
rule learning and rule identification. Although this study is indeed an investigation of the conditions under which subjects
acquire rule-governed behaviors, it seems
prudent at this stage to allow the term
implicit learning to maintain definitional
integrity apart from both rule learning and
rule identification, as those terms are traditionally used. The separation from rule
identification seems straightforward; the
process here is basically one of systematic
testing of existing hypotheses and rules. It
is an interesting problem but one very different from inquiring how those rules came
to be.
The argument for the nonsynonymity of
implicit learning and rule learning is subtler.
As a first approximation it is proposed that
rule learning subsumes at least two elementary processes: a primitive process of
apprehending structure by attending to frequency cues, and a more explicit process
whereby various mnemonics, heuristics, and
strategies are engaged to induce a representational system. The former is what is
defined here as implicit learning, and it has
certain conceptual similarities with the
differentiation
component of perceptual
learning (see Gibson, 1969). The latter
is what is listed in the psychological lexicon
under rule learning.
The paradigmatic confusion, however, is
not necessarily lessened by this classification. .In practice it is exceedingly difficult
to identify the boundary between rule learning and rule identification. I know of no
published report on rule learning in adults
where the rules to be learned were not immediate or trivial generalizations of well-practiced and readily retrievable rule systems.
Despite the nomenclature, cognitive psychologists rarely study rule learning.
The work of Miller and his associates on
artificial-grammar learning (Miller, 1967,
chap. 7) is illustrative. Their work, which
was begun using grammars of a complexity
approaching that of Figure 1 (Miller, 1958;
Shipstone, 1960), was shifted over to rela- ARTHUR S. REBER
tively simple systems on the grounds that
the synthetic languages used initially were
too intricate for "an afternoon in the laboratory." The outcome of "Project Grammarama" thus became, as Miller recognized,
an interesting procedure for evaluating the
process of explicit-rule induction. Essentially nothing was learned about the implicit
acquisition of highly complex systems like
languages. The problem was th…